Your product now passes through four filters before it reaches a person. All of them read the same catalog.

Prime Big Deal Days falls on October 6 and 7 this year [1]. For two days, millions of shoppers ask Alexa for Shopping what to buy, and it answers from whatever is in your listing at that moment.

That agent isn't a novelty anymore. Amazon reported that more than 350 million customers used Alexa for Shopping, the agent that replaced Rufus, in the last 12 months, with interactions up more than 5x year over year [2]. It already buys on its own when a product hits the price the customer set [3]. At Walmart, customers who shop with Sparky spend 40% more per order than those who don't [4].

Outside the marketplaces, the same thing is happening. In late September, NielsenIQ reported that 51% of US consumers used at least one AI tool to shop in the past month, the first time the number has passed the halfway mark [5]. Adobe measured a 393% year-over-year jump in AI-referred traffic to US retail sites in the first quarter of 2026. In March, that traffic converted 42% better than non-AI traffic and produced 37% more revenue per visit [6].

In my view, these numbers together are still being read as far too small. It's not that one more channel showed up. It's that new filters have moved in between your product and the person who's going to buy it. Today there are four.

Four filters between your product and the buyer

• •AI assistants. ChatGPT, Gemini, Claude, Perplexity. When someone asks what to buy, they build a product card out of whatever they can find about you: title, price, attributes, reviews, store. If the data isn't there, the card ships without it. Or it ships with your competitor.
• •Marketplace agents. Alexa for Shopping on Amazon, Sparky at Walmart. This is the most mature filter: it has the cart, the order history, the shipping promise and, on Amazon, it already completes the purchase without a click [3]. It reads what's in the listing: title, bullets, product attributes, Q&A, reviews. Only what's in the listing.
• •Personal agents. General-purpose assistants that browse sites, compare, and are starting to close the purchase on the consumer's behalf. In September, Meta launched Muse in the US, an agent that opens a browser, fills out forms and completes the checkout, and that you talk to in its own app or directly in WhatsApp [11]. Amazon already does something similar on third-party sites with Buy for Me [3], and Google launched UCP, an open protocol for agents to buy from any store, now powering checkout with Google Pay from Search and Gemini [7] [8]. It's the least mature layer, but the one that changes the consumer's relationship with buying the most: they delegate the whole task.
• •Your own site's search and assistant. LLM-powered search, recommendations, support that sells. It reads your catalog straight from the source, with every defect the source has.

Four filters, with different owners and different levels of maturity. What they have in common is what matters.

None of them is impressed by a nice photo

In the first piece I wrote on this blog, I argued that agentic commerce changes the logic of eligibility in e-commerce [9]. When a person lands on your listing, you can still persuade. When an agent evaluates your product, you first need to be understood.

That's how all four filters work. They read attributes, spec sheets, titles, images, and consistency across channels. An empty field is already a reason to skip you. A dimension that says one thing on Amazon and another on your site is a reason to distrust you.

Think about what happens when the bullets on your Amazon listing say the backpack fits a 15.6-inch laptop, the backend attribute says 14 inches, and your own site says 15. The agent has two paths. It can distrust you and recommend the competitor. Or it can recommend you and hand the customer something different from what it promised. In the first case, you vanish from the answer. In the second, you get a return.

Either way, the problem isn't media. It isn't ranking. It's consistency.

That's exactly what we see at GLOBALD. Across the more than 500,000 SKUs our compliance agent has analyzed, the most frequent errors aren't about persuasion. They're about consistency: dimensions that differ between image and text, colors that don't match, product names that change from one place to another, language errors, channel rule violations.

None of these problems takes your brand out of the conversation. All of them can take your product out of the purchase.

The strongest filter is inside the marketplace

There's a common reading that AI is going to pull product search out of Amazon. I don't think that's quite it.

In the US, 57% of consumers start their product search on Amazon, against 42% on search engines [10]. That hasn't changed. AI-referred traffic is growing fast, but even after a 393% jump it's still a small slice of total retail visits.

What changes isn't where the customer searches. It's what answers the search. Instead of a results page sorted by rank and sponsored placements, an agent that reads the listing, compares attributes, and picks. Alexa for Shopping doesn't look at your main image the way a shopper does. It reads the title, the bullets, the product attributes, the Q&A, and the reviews, and it answers the shopper's question from that. If the bullets say one thing and the attributes say another, it has to decide which one to believe. Amazon already does this with 350 million customers. Walmart's Sparky user base grew 70% in a year [4].

Prime Big Deal Days is the first live test of the season. Whatever the agent reads during those two days is what you shipped the week before.

For sellers, this means your product now gets filtered twice. By the agent outside, which decides whether you make it into the answer. And by the agent inside, which decides among the listings that survived. Both read the same thing.

The good part: the base is the same for all four

What the four filters require at the base is identical: attributes that are complete, correct, and consistent with each other. That work is done once, at the source.

The output isn't. Each filter reads the product in its own format. Amazon has its own attribute schema and title rules. Walmart Marketplace has its own spec. AI assistants read feeds and structured data on the page. UCP has its standard. Your site has its schema. The same product has to be translated for each one.

The difference between having the base and not having it is what happens in that translation. Without the base, every channel becomes a rewrite, and every rewrite carries a different error. That's how the laptop size ends up different in the bullet, in the backend attribute, and on your site. With the base, translation becomes mapping: an attribute fixed at the source comes out right in the ChatGPT card, the Amazon listing, the personal agent's purchase, and your site's search.

That doesn't mean it's little work. Fixing a few dozen products is simple. Keeping thousands of SKUs consistent, in more than one language, translated for channels with different rules, and updated every week, is not. That's why this layer rarely gets solved with a spreadsheet and good intentions. It needs to become infrastructure: one source of truth and several outputs, each in the channel's format.

The catalog, long treated as back office, has become the interface between your product and the agents that decide on its behalf.

Where to start before Black Friday

Prime Big Deal Days is the dress rehearsal. Black Friday is November 27 and Cyber Monday is November 30. Less than eight weeks out, there's time to do the essentials:

• •Ask the agents before your customer does. ChatGPT, Gemini, and Alexa for Shopping. Ask the questions a buyer in your category would ask, with a real need and a real budget. See whether your product makes the list, with what information, and where the link goes.
• •Compare what the agent says with what the listing says. Every discrepancy in dimensions, capacity, color, or compatibility between bullets, backend attributes, and your own site is a lost sale or a return waiting to happen.
• •Keep a source of truth outside the marketplace. The listing can't be the only place your product is well described. The same data needs to feed Amazon, Walmart, your own site, and the feeds the AIs read.
• •Write attributes that answer questions. The agent decides on what the buyer asked. If they want battery life and your listing doesn't say, you're not in the conversation.
• •Start with what sells most. You can't fix everything by November. You can fix what weighs most.

The catalog the agents will read

For years, the question sellers asked was: does my page convert?

It's still valid. But with four agents in the way, the first question has become a different one:

If the answer is no, no amount of media or ranking investment fixes it. The agent has already picked someone else.

At GLOBALD, this is exactly the kind of problem that interests us: making the product understood by every agent that decides on its behalf. We audit the catalog, find the compliance, attribute, text, and image errors that keep the product out of the answers, fix them, and distribute the data, in each channel's format, to marketplaces, your own site, LLMs, and agent protocols like UCP.

Black Friday is November 27. The catalog the agents will read that day is whatever is live over the next few weeks.